Robust feature learning using contractive autoencoders for multi-omics clustering in cancer subtyping.

Area of Science:

• Computational Biology
• Bioinformatics
• Cancer Research

Background:

• Precise cancer subtyping is crucial for diagnosis, treatment, and prognosis.
• Integrating multi-omics data offers potential for improved cancer subtyping.
• Challenges exist in handling data heterogeneity and extracting relevant features for multi-omics integration.

Purpose of the Study:

• To develop a novel multi-omics clustering framework for enhanced cancer subtyping.
• To address the challenge of data heterogeneity in multi-omics integration.
• To improve the accuracy of cancer patient stratification by incorporating survival information.

Main Methods:

• Utilized contractive autoencoder (CAE) for robust feature extraction from multi-omics data.
• Incorporated Cox proportional hazards regression to select survival-associated features.
• Applied K-means clustering on integrated features for cancer subtyping.

Main Results:

• The proposed framework effectively integrated four types of omics data across ten cancer datasets.
• The method demonstrated superior performance compared to existing approaches, evidenced by higher C-index scores.
• The framework achieved more significant differences in survival curves, indicating improved patient stratification.

Conclusions:

• The developed multi-omics clustering framework provides an effective approach for cancer subtyping.
• The integration of robust features and survival information enhances diagnostic and prognostic accuracy.
• Further analyses, including differential gene and pathway enrichment, support the framework's effectiveness.